摘要
针对变电站入侵异物尺度差异大和模型复杂的问题,提出了一种基于多尺度线性注意力机制与YOLOv5融合的变电站异物入侵检测算法.首先,采用轻量级多尺度线性注意力模型(EfficientViT)作为主干网络,提升模型多尺度学习能力.其次,将颈部网络三卷积跨阶段局部瓶颈模块(C 3)中的瓶颈层替换为轻量级的幻影卷积瓶颈层,降低模型的复杂度.最后,设计了三重增强注意力机制(TEA),并将其融入颈部网络C3模块中,通过各维度之间的交互和增大的感受野,有效地融合特征信息,提高检测精度.实验结果显示,与原模型相比,改进模型平均精度均值提高了 1.4％,浮点运算次数降低了 23.9％,提高异物检测精度的同时降低了模型复杂度,有利于实际中的应用.
Abstract
To address the problem of large-scale differences and complex models of foreign objects intruding into substations,a substation foreign object intrusion detection algorithm is proposed based on the fusion of multi-scale linear attention mechanism and YOLOv5.First,a lightweight multi-scale linear attention model(EfficientViT)is employed as the backbone network to improve the multi-scale learning ability of the model.Second,the bottleneck layer in the triple convolution cross-stage local bottleneck module(C3)of the neck network is replaced with a lightweight phantom convolution bottleneck layer to reduce the complexity of the model.Finally,a triple-enhanced attention mechanism(TEA)is designed and integrated into the C3 module of the neck network.The feature information is effectively fused to improve the detection accuracy by the interaction between dimensions and the enlarged receptive fields.Experimental results show that compared with the original model,the average precision of the improved model is increased by 1.4％,and the number of floating-point operations is reduced by 23.9％.The accuracy of foreign object detection is improved while the complexity of the model is reduced,which is conducive to practical application.
NETL
NSTL
万方数据